158 lines
5.4 KiB
JSON
158 lines
5.4 KiB
JSON
{
|
|
"benchmark_manifest_sha256": "0ea5ab07f46c509a7a24943b31d5e9bfd6368920609227bc47613e18e52c4642",
|
|
"critical_subgroup_policy": "Any required subgroup with insufficient support, missing metrics, a failed non-inferiority comparison or regression blocks promotion; averages cannot override it.",
|
|
"decision": "At least one mandatory Phase 4 gate failed; Phase 5 remains blocked.",
|
|
"gate_policy": "geointel-p4-evaluation-harness-v2",
|
|
"invalid_gate_states": {},
|
|
"local_gates": {
|
|
"all_declared_evaluator_families_exercised": {
|
|
"declared": [
|
|
"change_detection",
|
|
"footprint_segmentation",
|
|
"geospatial_data_validation",
|
|
"object_detection",
|
|
"raster_classification",
|
|
"terrain_interpretation",
|
|
"vector_comparison"
|
|
],
|
|
"observed": [
|
|
"change_detection",
|
|
"footprint_segmentation",
|
|
"geospatial_data_validation",
|
|
"object_detection",
|
|
"raster_classification",
|
|
"terrain_interpretation",
|
|
"vector_comparison"
|
|
],
|
|
"status": "pass"
|
|
},
|
|
"complete_raw_predictions_retained": {
|
|
"status": "pass",
|
|
"violating_samples": []
|
|
},
|
|
"implemented_capability_inventory": {
|
|
"capability_count": 15,
|
|
"status": "pass"
|
|
},
|
|
"manifest_training_firewall_contract": {
|
|
"checks": {
|
|
"clean_train_allowed": true,
|
|
"non_train_role_blocked": true,
|
|
"protected_path_blocked": true,
|
|
"renamed_protected_lineage_blocked": true
|
|
},
|
|
"status": "pass"
|
|
},
|
|
"normative_split_roles_and_leakage": {
|
|
"leakage_status": "pass",
|
|
"observed_roles": [
|
|
"background-test",
|
|
"calibration",
|
|
"challenge",
|
|
"test",
|
|
"train",
|
|
"val"
|
|
],
|
|
"required_roles": [
|
|
"background-test",
|
|
"calibration",
|
|
"challenge",
|
|
"test",
|
|
"train",
|
|
"val"
|
|
],
|
|
"status": "pass"
|
|
},
|
|
"protected_operating_point_contract": {
|
|
"evidence": "Protected cases carry pre-registered configurations. Fixed AP/risk-coverage diagnostics cannot select an operating point or feed back into training.",
|
|
"status": "pass"
|
|
},
|
|
"reference_implementation_baseline": {
|
|
"status": "pass"
|
|
},
|
|
"stratified_metric_contract": {
|
|
"observed_overall_status": "not_evaluable",
|
|
"status": "pass"
|
|
},
|
|
"undefined_metric_truth_table": {
|
|
"sample_id": "background-test-pure-empty",
|
|
"status": "pass"
|
|
}
|
|
},
|
|
"local_harness_status": "pass",
|
|
"missing_gate_names": {
|
|
"local": [],
|
|
"product": []
|
|
},
|
|
"numeric_model_release_targets": "not_frozen_without_reviewed_representative_incumbent_baseline",
|
|
"phase_decision": "blocked",
|
|
"product_benchmark_status": "fail",
|
|
"product_gates": {
|
|
"active_model_available_and_hash_verified": {
|
|
"configured_path": "/app/models/geointel-building-yolov8s-smallbld-minpx3-img640-ft30.pt",
|
|
"configured_sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1",
|
|
"reason": "Configured active model is not locally accessible.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"authoritative_reference_portfolio_available": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"executed_product_incumbent_baseline": {
|
|
"checked_artifacts": [],
|
|
"derived_gates": {
|
|
"authoritative_reference_portfolio_available": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"human_review_complete": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"protected_storage_isolation": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"representative_product_subgroup_support": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"split_independence": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
}
|
|
},
|
|
"expected_path": "artifacts/evidence/accuracy/P4/product-baseline-manifest.json",
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"human_review_complete": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"phase3_leakage_resolved": {
|
|
"observed": "attention",
|
|
"status": "fail"
|
|
},
|
|
"protected_storage_isolation": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"representative_product_subgroup_support": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
},
|
|
"split_independence": {
|
|
"reason": "No executed, hash-bound product incumbent baseline manifest is available.",
|
|
"status": "not_evaluable"
|
|
}
|
|
},
|
|
"promotion_allowed": false,
|
|
"schema_version": 2,
|
|
"status": "fail",
|
|
"unexpected_gate_names": {
|
|
"local": [],
|
|
"product": []
|
|
}
|
|
}
|